Quantum Learning Pathway

Quantum Learning Pathway

Created: 2026-06-30
Derived from: QUANTUM_BASELINE_2026-06-30.md (what you have) +
quantum_rag_demo/docs/research/expansion_2026-06-30/ (what the literature says).
Goal: a learn-by-evidence path — spend effort where the literature says there's real
ground, learn-by-building where it's actionable, and read-only where it's speculative.

Guiding principle

The research expansion stratified your 7 retrieval methods by evidence. Let that drive how
you learn each one:

Everything below maps to files you already have, so learning and project progress are the same activity.

Prerequisites (you already have these)

• ✅ Working quantum_env (Py 3.13, qiskit 2.4.2) — verified 2026-06-30.
• ✅ Notebooks quantum_computing/01_quantum_fundamentals.ipynb and 02_grover_for_vector_search.ipynb — start here if rusty on qubits/gates/Grover.
• ✅ Reference implementations of all 7 methods in enhanced_quantum_rag.py.

Stage 0 — Stabilize (do first; unblocks everything)

Learning goal: understand the codebase by making it green and modern.

1. Fix the 2 failing tests — research_agent/test_token_optimizer.py (token-reduction
consistency). Likely tokenizer/version drift. Target: 583/583 green.
2. Qiskit 3.0 readiness — the suite throws many 2.x deprecation warnings
(GroverOperator, MCXGate.get_num_ancilla_qubits, mcx(mode=...), execute()).
Learn the new primitives: SamplerV2/EstimatorV2, BackendV2, target-based transpile.
Read: expansion report §6.1.
3. Pin requirements.txt — add qiskit-algorithms, pytest, pytest-asyncio,
scikit-learn, matplotlib (all needed at baseline but unpinned).

Done when: 583/583 tests pass with zero deprecation warnings.

Stage 1 — Tier-1 build (highest value)

Learning goal: the two things the literature actually validates.

1a. ML error mitigation (the single best-supported result)

• Concept: a CNN autoencoder learns to denoise NISQ outputs — fidelity 0.298→0.774 on

5-qubit circuits, no extra qubits. (Refs [27],[31] in the report; corpus
combined_ml_quantum_error_correction.csv.)

• Build: wire it into your existing Grover circuits as a post-processing layer. This is

exactly what ai_extensions/error_predictor.py gestures at — so *this also validates
ai_extensions* (currently zero runtime evidence).

• Learn: density-matrix reconstruction, synthetic noise models (depolarizing, amplitude/

phase damping), fidelity-aware loss.

• Success metric: fidelity Δ ≥ 0.3 on a 5-qubit test circuit.

1b. Hybrid retrieval pipeline (classical filter → quantum refine)

• Concept: BM25/FAISS reduces N→K, Grover refines on K. Preserves √K speedup, sidesteps

the QRAM loading wall. The only deployable quantum-retrieval architecture today.
(Refs [8],[26]; QRAG reports 40–50% latency cut.)

• Build: you already have the pieces — mvp_rag.py (FAISS) + enhanced_quantum_rag.py

(Grover). Formalize the K-candidate handoff and measure.

• Learn: why QRAM kills the unstructured case; where √K actually helps.
• Success metric: measurable improvement on K=100 vs classical-only rerank.

Stage 2 — Tier-2 prototype + benchmark

Learning goal: produce the head-to-head retrieval benchmarks the literature is missing —
this is where you could publish.

The recurring gap across the corpus: *no one benchmarks these against BM25/FAISS on BEIR/
MS MARCO.* Closing that gap is both the learning exercise and a first-mover contribution.

• Complex/phase embeddings (enhanced_quantum_rag.py ComplexEmbeddingEncoder) → run

on BEIR vs BM25. Read: Semantic Wave Functions [5], Harvey [7], Phase-Coded RAG [8].

• HAM / VSA (hyperdimensional.py, graphd.py) → controlled recall/latency vs HNSW at

100K. Learn the capacity–accuracy tradeoff (accuracy degrades with superposition load).
Read: krop embeddings [17], Memory Tesseract [19].

• Grover for structured data → learn the FRANS result [9]: state-prep drops to

O(poly log N) only for geometrically structured embedding spaces. That's the regime where
Grover-KNN is genuinely viable.

Bonus / classically-implementable (no quantum HW):

• HD-NSW binary retrieval [4] — 8× memory cut, >860 QPS @ 10M docs. Well-characterized

on BEIR; great learning project.

• Phase-Coded Memory RAG [8] — frequency-domain retrieval, runs on classical hardware.

Stage 3 — Tier-3 study (read-only until a gate passes)

Learning goal: know the theory; resist sinking build time into the unvalidated.

• Quantum-kernel attention — your biggest unvalidated claim; **no retrieval experiment

exists in the entire corpus.** Study quantum kernels conceptually; gate: only build if an
expanded search turns up an empirical result.

• Quantum-KNN at production scale — QRAM-blocked. Study SWAP-test KNN [20] (89–96% on

Iris/MNIST at tiny QRAM) to understand why it doesn't scale.

• Compositional QNLP (DisCoCirc) [34] — trapped-ion QA that beat GPT-4/LSTM on a

compositional task; QRAM-free. Highest-ceiling frontier topic; read now, build only with
hardware/simulator access.

Suggested cadence

Decision gates (from the report)

• Wk 4: if ai_extensions characterization shows >50% components have no clear purpose → refactor before validating.
• Wk 10: if expanded quantum-kernel-attention search still finds no empirical result → deprioritize/remove it.
• Wk 14: if CNN mitigation gain <0.2 fidelity → switch to GNN/transformer mitigation.

Pointers

• What you have: QUANTUM_BASELINE_2026-06-30.md
• Why these priorities: quantum_rag_demo/docs/research/expansion_2026-06-30/research_expansion_report.md
• Paper corpus to mine: .../expansion_2026-06-30/corpus/ (66 CSVs, ~7,195 ranked rows)
• Fundamentals refresher: quantum_computing/01_.ipynb, 02_.ipynb